Code generation for JSON Schema (Draft 07).

New in version 0.87 – the generator now recognises a special case of anyOf or oneOf to specify nullability. See below.

New in version 0.84 – the generator will now recognise the not schema for most validations, and will output reversed validation checks. For example, "not": { "const": "error" } in a property sub-schema will test that a string is not equal to "error".

Added to the code generator – the ability to configure generation options using a JSON or YAML file. See the documentation at CONFIG.md.

Also build tool support – see below.

Also, the ability to add annotations to generated classes – see annotations.

And from version 0.86 onward, the ability to force the output of a companion object for all or selected classes – see companionObject.

JSON Schema provides a means of describing JSON values – the properties of an object, constraints on values etc. – in considerable detail. Many APIs now use JSON Schema to specify the content of JSON parameters and response objects, either directly or as part of an OpenAPI specification, and one way of ensuring conformity to the specified schema is to generate code directly from the schema itself.

This is not always possible – some characteristics described by a schema may not be representable in the implementation language. But for a large subset of schema definitions a viable code representation is achievable, and this library attempts to provide conversions for the broadest possible range of JSON Schema definitions.

The library uses a template mechanism (employing Mustache templates), and templates are provided to generate classes in Kotlin and Java, or interfaces in TypeScript.

Simply create a CodeGenerator, supply it with details like destination directory and package name, and invoke the generation process:

        val codeGenerator = CodeGenerator()
        codeGenerator.baseDirectoryName = "output/directory"
        codeGenerator.basePackageName = "com.example"
        codeGenerator.generate(File("/path/to/example.schema.json"))

The resulting code will be something like this (assuming the schema is the one referred to in the introduction to json-kotlin-schema):

package com.example

import java.math.BigDecimal

data class Test(
    /** Product identifier */
    val id: BigDecimal,
    /** Name of the product */
    val name: String,
    val price: BigDecimal,
    val tags: List<String>? = null,
    val stock: Stock? = null
) {

    init {
        require(price >= cg_dec0) { "price < minimum 0 - $price" }
    }

    data class Stock(
        val warehouse: BigDecimal? = null,
        val retail: BigDecimal? = null
    )

    companion object {
        private val cg_dec0 = BigDecimal.ZERO
    }

}

Some points to note:

• the generated class is an immutable value object (in Java, getters are generated but not setters)
• validations in the JSON Schema become initialisation checks in Kotlin
• nested objects are converted to Kotlin nested classes (or Java static nested classes)
• fields of type number are implemented as BigDecimal (there is insufficient information in the schema in this case to allow the field to be considered an Int or Long)
• non-required fields may be nullable and may be omitted from the constructor (the inclusion of null in the type array will allow a field to be nullable, but not to be omitted from the constructor)
• a description will be converted to a KDoc comment in the generated code if available

The Code Generator can process a single file or multiple files in one invocation. The generate() function takes either a List or a vararg parameter array, and each item may be a file or a directory. In the latter case all files in the directory with filenames ending .json or .yaml (or .yml) will be processed.

It is preferable to process multiple files in this way because the code generator can create references to other classes that it knows about – that is, classes generated in the same run. For example, if a properties entry consists of a $ref pointing to a schema that is in the list of files to be generated, then a reference to an object of that type will be generated (instead of a nested class).

It is important to note that the output code will be "clean" – that is, it will not contain annotations or other references to external libraries. I recommend the use of the kjson library for JSON serialisation and deserialisation, but the classes generated by this library should be capable of being processed by the library of your choice.

There is one exception to this – classes containing properties subject to "format" validations will in some cases (e.g. "email", "ipv4") cause references to the json-validation library to be generated, and that library must be included in the build of the generated code.

The default in the JSON Schema specification for additionalProperties is true, meaning that any additional properties in an object will be accepted without validation. Many schema designers will be happy with this default, or will even explicitly specify true, so that future extensions to the schema will not cause existing uses to have problems.

Unfortunately, there is no straightforward implementation for additionalProperties in code generation, so the setting will be taken as false even if it is specified otherwise. Most JSON deserialisation libraries have a means of specifying that additional properties are to be ignored; for kjson the allowExtra variable (Boolean) in JSONConfig must be set to true.

The code generator will create a data class whenever possible. This has a number of advantages, including the automatic provision of equals and hashCode functions, keeping the generated code as concise and readable as possible.

Unfortunately, it is not always possible to use a data class. When the generated code involves inheritance, with one class extending another, the base class will be generated as an open class and the derived class as a plain class.

In these cases the code generator will supply the missing functions – equals, hashCode, toString, copy and the component[n] functions that would otherwise be created automatically for a data class.

The standard way of specifying a value as nullable in JSON Schema is to use the type keyword:

    { "type": [ "object", "null" ] }

When the code generator encounters a property or array item defined in this way, it will make the generated type nullable.

A problem arises when we consider the interaction of this declaration of nullability with the required keyword. What should the generator produce for an object property that does not include null as a possible type, but is not in the required list? The solution adopted by the code generator is to treat the property as if it had been defined to allow null, and this seems to work well for the majority of cases, although strictly speaking, it is not an accurate reflection of the schema.

In particular, it helps with the case of utility sub-schema which is included by means of $ref in multiple places, in some cases nullable and in some cases not. For example, an invoice may have a billing address and an optional delivery address, both of which follow a common pattern defined in its own schema. The shared definition will have "type": "object", but the delivery address will need to be nullable, so generating a nullable type for a reference omitted from the required list will have the desired effect.

But this solution does not work for all circumstances. For example, it does not cover the case of an included sub-schema as an array item – there is no required for array items.

One way of specifying such a schema using the full capabilities of JSON Schema is as follows:

{
  "type": "object",
  "properties": {
    "billingAddress": {
      "$ref": "http://example.com/schema/address"
    },
    "deliveryAddress": {
      "anyOf": [
        { "$ref": "http://example.com/schema/address" },
        { "type": "null" }
      ]
    }
  },
  "required": [ "billingAddress", "deliveryAddress" ]
}

It is not easy to generate code for the general case of oneOf or anyOf, but the code generator will detect this specific case to output the deliveryAddress as nullable:

1. The anyOf or oneOf array must have exactly two sub-schema items
2. One of the items must be just { "type": "null" }

In this case, the code generator will generate code for the other sub-schema item (the one that is not { "type": "null" }, often a $ref), and treat the result as nullable.

(NOTE – the configuration file may be a simpler way to specify custom classes, particularly when combined with other configuration options. See the Configuration Guide.)

The code generator can use custom types for properties and array items. This can be valuable when, for example, an organisation has its own custom classes for what are sometimes called "domain primitives" – value objects representing a fundamental concept for the functional area.

A common example of a domain primitive is a class to hold a money value, taking a String in its constructor and storing the value as either a Long of cents or a BigDecimal.

There are three ways of specifying a custom class to the code generator:

1. URI
2. Custom format types
3. Custom keywords

An individual item in a schema may be nominated by the URI of the element itself. For example, in the schema mentioned in the Quick Start section there is a field named price. To specify that the code generator is to use a Money class for this field, use:

        codeGenerator.addCustomClassByURI(URI("http://pwall.net/test#/properties/price"), "com.example.Money")

The base URI can be either the URL used to locate the schema file or the URI in the $id of the schema.

A distinct advantage of this technique is that when a $ref is used to share a common definition of a field type, the destination of the $ref can be specified to the code generator function shown above, and all references to it will use the nominated class. It is also the least obtrusive approach – it does not require modification to the schema or non-standard syntax.

The JSON Schema specification allows for non-standard format types. For example, if the specification of the property in the schema contained "format": "x-local-money", then the following will cause the property to use a custom class:

        codeGenerator.addCustomClassByFormat("x-local-money", "com.example.Money")

The JSON Schema specification also allows for completely non-standard keywords. For example, the schema could contain "x-local-type": "money", in which case the following would invoke the use of the custom class:

        codeGenerator.addCustomClassByExtension("x-local-type", "money", "com.example.Money")

This code generator targets the Draft-07 of the JSON Schema specification, and it includes some features from Draft 2019-09.

It also includes support for the int32 and int64 format types from the OpenAPI 3.0 Specification.

A CodeGenerator object is used to perform the generation. It takes a number of parameters, many of which can be specified either as constructor parameters or by modifying variables in the constructed instance.

• targetLanguage – a TargetLanguage enum specifying the target language for code generation (the options are KOTLIN, JAVA or TYPESCRIPT – TypeScript coverage is not as advanced as that of the others at this time)
• templateName – the primary template to use for the generation of a class
• enumTemplateName – the primary template to use for the generation of an enum
• basePackageName – the base package name for the generated classes (if directories are supplied to the generate() function, the subdirectory names are used as sub-package names)
• baseDirectoryName – the base directory to use for generated output (in line with the Java convention, output directory structure will follow the package structure)
• derivePackageFromStructure – a boolean flag (default true) to indicate that generated code for schema files in subdirectories are to be output to sub-packages following the same structure
• generatorComment – a comment to add to the header of generated files
• markerInterface – a “marker” interface to be added to every class

The configure() function takes a File or Path specifying a configuration file. See CONFIG.md for details of the configuration options.

There are two generate() functions, one taking a List of Files, the other taking a vararg list of File arguments. As described above, it is helpful to supply all the schema objects to be generated in a single operation.

While the generate() functions take a file or files and convert them to an internal form before generating code, the generateClass() and generateClasses() functions take pre-parsed schema objects. This can be valuable in cases like an OpenAPI file which contains a set of schema definitions embedded in another file.

The generateAll() function allows the use of a composite file such as an OpenAPI file containing several schema definitions. For example, an OpenAPI file will typically have a components section which contains definitions of the objects input to or output from the API. Using the generateAll() function, the set of definitions can be selected (and optionally filtered) and the classes generated for each of them.

To simplify the use of the code generator in conjunction with the common build tools the following plugins will perform code generation as a pre-pass to the build of a project, allowing classes to be generated and compiled in a single operation:

• json-kotlin-gradle
• json-kotlin-maven

The latest version of the library is 0.87, and it may be obtained from the Maven Central repository.

    <dependency>
      <groupId>net.pwall.json</groupId>
      <artifactId>json-kotlin-schema-codegen</artifactId>
      <version>0.87</version>
    </dependency>

    implementation 'net.pwall.json:json-kotlin-schema-codegen:0.87'

    implementation("net.pwall.json:json-kotlin-schema-codegen:0.87")

Peter Wall

2023-02-12